A Geometry-Aware AI Emulator for the Coupled Whole Atmosphere from Earth Surface to the Ionosphere and Thermosphere

Whole-atmosphere models such as WACCM-X resolve coupling from the Earth surface to the Mesosphere-Lower-Thermosphere (MLT), and Ionosphere-Thermosphere (IT) systems with expensive computational costs. Here we introduce CAM-NET, a geometry-aware Spherical Fourier Neural Operator (SFNO) surrogate for emulating WACCM-X variability from Earth surface to IT region. CAM-NET is trained on 3-hourly WACCM-X simulations and predicts neutral winds, temperature, pressure-coordinate vertical velocity, electron density, and zonal ion drift. The framework combines a Spherical Fourier Neural Operator (SFNO) backbone with a newly developed lightweight module that extends the frozen atmospheric representation to plasma variables. For the held-out simulation, CAM-NET preserves the dominant IT morphology and remains stable during multi-day autoregressive rollouts. Spherical-harmonic diagnostics show that the model retains low-degree variability while damping high-wavenumber mesospheric structures, especially near 90 km where gravity wave breaks. CAM-NET is intended as a computationally efficient emulator of WACCM-X, rather than an operational forecasting system. These results demonstrate its potential for rapid ensemble experiments, uncertainty quantification, and sensitivity studies of large-scale coupled whole atmospheric variability.